Insulator Self-Explosion Defect Detection Based on Multi-Task Learning Convolutional Neural Network

Abstract

Unmanned Aerial Vehicles (UAVs) have been used for power transmission line inspections, and
detecting insulator defects from the massive data obtained from aerial photography has become
a hot issue. Most of the existing algorithms first locate the insulator and then detect the defect
position, requiring training two networks. In this paper, the problem of self-explosion defect
detection is converted into the classification problem of picture blocks, and an insulator
self-explosion defect detection model based on multi-task learning convolution neural network
is proposed. To address the scarcity of defect images, in addition to traditional data
augmentation methods, synthetic pictures and digitally-altered pictures are used for data
augmentation. Experiment results show that the new data augmentation method can effectively
increase classification accuracy. The precision rate and the racall rate of defect detection of the
proposed model are 0.989 and 0.942, indicating the method meets the requirements for insulator
defect detection.